NanoClaw

NanoClaw is a small, self-hosted personal AI agent built around a Node.js host process and isolated agent containers. Its compact codebase is designed for operators who want to inspect and customize the assistant rather than enable every channel and provider in the core. Messaging adapters and related capabilities are installed as skills, while channel routing can direct different accounts or conversations to separate agents or a shared agent. The canonical repository publishes NanoClaw under the MIT license.

Isolation is central to the design. Agent groups run in containers with their own workspaces, memory, and permitted mounts; the project also documents stronger platform-specific isolation options. Credential handling can use OneCLI Agent Vault so an agent does not need to receive raw provider keys directly. NanoClaw itself is source-available without a hosted subscription, but running it requires local or server infrastructure plus any selected model, messaging, or vault services. Those external services may have their own fees and terms.

Top features

• Compact host architecture: a single Node.js process handles channel events and launches agent work.
• Per-agent containers: isolate groups, workspaces, memory, and filesystem mounts so unrelated assistant contexts do not automatically share data.
• Skill-installed channels: add only the messaging or service adapters needed for a deployment instead of enabling a large default integration set.
• Flexible routing: keep work and personal channels separate, map several channels to one assistant, or assign a dedicated agent to a sensitive workflow.
• Credential mediation: documented OneCLI Agent Vault integration injects credentials into approved outbound requests rather than exposing raw keys to the agent.
• User-owned customization: operators can change the source and skill set to match their own assistant behavior and deployment policies.

Use cases

• Run a self-hosted assistant reachable from one or more supported messaging systems.
• Separate personal, work, and experimental assistants into different containers and storage scopes.
• Audit and modify a smaller codebase before trusting it with files, tools, or communication channels.
• Add a channel adapter only when it is needed and review that skill before enabling it.
• Keep model credentials behind an external vault while allowing narrowly approved provider requests.